Issue2458

Tornado should run on Jython, and it does not appear to really rely on fcntl (#1943). Instead, Tornado mainly seems to be missing some functionality in our ssl implementation, likely minor.

Apply the following diff to a clone of https://github.com/tornadoweb/tornado:

$ git diff
diff --git a/tornado/platform/auto.py b/tornado/platform/auto.py
index 449b634..6d0f435 100644
--- a/tornado/platform/auto.py
+++ b/tornado/platform/auto.py
@@ -35,6 +35,10 @@ if 'APPENGINE_RUNTIME' in os.environ:
 elif os.name == 'nt':
     from tornado.platform.common import Waker
     from tornado.platform.windows import set_close_exec
+elif os.name == 'java':
+    from tornado.platform.common import Waker
+    def set_close_exec(fd):
+        pass
 else:
     from tornado.platform.posix import set_close_exec, Waker

set_close_exec can be a no-op on Jython because Java opens files such that the file descriptors are closed on exec; see http://stackoverflow.com/a/9224348/423006

After installing dependencies (backports-abc, singledispatch), $ jython -m tornado.test.runtests. But see its tox.ini for more options. Many tests curerently pass, but Tornado is failing in some aspects of using SSL, such as missing attributes on SSL sockets and possibly exception handling.

One more change is required to run Tornado's hello, world application (http://www.tornadoweb.org/en/stable/#hello-world), due to #2406:

$ git diff tornado/netutil.py
diff --git a/tornado/netutil.py b/tornado/netutil.py
index b249945..5dbd94e 100644
--- a/tornado/netutil.py
+++ b/tornado/netutil.py
@@ -181,7 +181,7 @@ def bind_sockets(port, address=None, family=socket.AF_UNSPEC,
             # ipv6 sockets and use a separate ipv4 socket when needed.
             #
             # Python 2.x on windows doesn't have IPPROTO_IPV6.
-            if hasattr(socket, "IPPROTO_IPV6"):
+            if hasattr(socket, "IPPROTO_IPV6") and os.name != "java":
                 sock.setsockopt(socket.IPPROTO_IPV6, socket.IPV6_V6ONLY, 1)

         # automatic port allocation with port=None

With Apache Benchmark using ab -k -c 20 -n 50000 localhost:8888/

Percentage of the requests served within a certain time (ms)
  50%     16
  66%     17
  75%     21
  80%     22
  90%     28
  95%     42
  98%     48
  99%     54
 100%    247 (longest request)

Not the fastest event loop (see Fireside running on Jetty with https://github.com/jimbaker/hellowsgi), but quite likely it can be made much faster if the event loop was based on Netty instead of the emulation with socket/ssl/select.

Tornado on CPython, same Apache Benchmark settings:

  50%      8
  66%      8
  75%      8
  80%      8
  90%      8
  95%      8
  98%      8
  99%      9
 100%     26 (longest request)

Still not as fast as Fireside, but significantly less overhead than Tornado/Jython currently. Of course this is timing the event loop - we are only doing small requests.

Is this in the trunk or just a private experiment? That is, should it be closed? Is #2604 formally a dependency?

If we have a bunch of SSL and IPv6 things to deal with, then is suggest it is for beyond 2.7.2.

@Jeff, consider it an experiment on the compatibility of our socket/select/ssl implementation by using an established framework.

There is much higher priority work, so I wouldn't target for 2.7.2